Super Agent Presentation v0.3.0

Highlights

• Adds a standard-library-only preregistered study controller for immutable plans, paired generation records, blinding, rating freeze, aggregation, and fail-closed claim gates.
• Adds a typed Codex host adapter with inert planning, explicit execution authorization, fixed shell-free argv, executable/workspace/Skill receipts, bounded captures, and honest unsupported-control reporting.
• Separates caller and controller-stored generation schemas, adds portable generated-image ingest, and binds records, responses, transcripts, host receipts, assignment keys, blind packets, and ratings against drift.
• Adds public-design gates for held-out coverage, external per-unit isolation, multiple model revisions and repeats, context/compaction observations, final audit, display checks, visual precision/recall, and semantic-slot density per 1,000 output tokens.
• Removes the redundant automatic conclusions module from experiment reports; the default experiment bundle is now 11,597 characters while retaining calibrated conclusion guidance.

Evidence and claim boundary

The checked-in Codex pilot is a one-case, one-repeat integration and activation observation. It remains permanently marked insufficient_evidence and does not establish quality, readability, or efficiency improvement.

Release verification passed on Python 3.9, 3.12, and 3.14: 263 unit tests per matrix job, fixture smoke, route-bundle smoke, and byte-for-byte distribution reproducibility. Independent review found no release blockers.

See CHANGELOG.md, BENCHMARK.md, and docs/adr/ADR-006-preregistered-study-pipeline.md for the complete protocol and limitations.